1 / 46

15-853:Algorithms in the Real World

15-853:Algorithms in the Real World. Indexing and Searching I (how google and the likes work). Indexing and Searching Outline. Introduction: model, query types Inverted File Indices: Compression, Lexicon, Merging Vector Models: Latent Semantic Indexing:

vernados
Download Presentation

15-853:Algorithms in the Real World

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. 15-853:Algorithms in the Real World Indexing and Searching I (how google and the likes work) 15-853

  2. Indexing and Searching Outline • Introduction: model, query types • Inverted File Indices: Compression, Lexicon, Merging • Vector Models: • Latent Semantic Indexing: • Link Analysis: PageRank (Google), HITS • Duplicate Removal: 15-853

  3. Indexing and Searching Outline • Introduction: • model • query types • common techniques (stop words, stemming, …) • Inverted File Indices: Compression, Lexicon, Merging • Vector Models: • Latent Semantic Indexing: • Link Analysis: PageRank (Google), HITS • Duplicate Removal: 15-853

  4. Basic Model • Applications: • Web, mail and dictionary searches • Law and patent searches • Information filtering (e.g., NYT articles) • Goal: Speed, Space, Accuracy, Dynamic Updates “Document Collection” Index Query Document List 15-853

  5. How big is an Index? • Sep 2003, self proclaimed sizes (gg = google, atw = alltheweb, ink = inktomi, tma = teoma) • Source: Search Engine Watch Billion Pages 15-853

  6. Sizes over time 15-853

  7. Precision and Recall • Typically a tradeoff between the two. number retrieved that are relevant total number retrieved Precision: number relevant that are retrieved total number relevant Recall: 15-853

  8. Precision and Recall • Does the black or the blue circle have higher precision? 15-853

  9. Main Approaches • Full Text Searching • e.g. grep, agrep (used by many mailers) • Inverted File Indices • good for short queries • used by most search engines • Signature Files • good for longer queries with many terms • Vector Space Models • good for better accuracy • used in clustering, SVD, … 15-853

  10. Queries • Types of Queries on Multiple “terms” • boolean (and, or, not, andnot) • proximity (adj, within <n>) • keyword sets • in relation to other documents • And within each term • prefix matches • wildcards • edit distance bounds 15-853

  11. Technique used Across Methods • Case folding London -> london • Stemming compress = compression = compressed (several off-the-shelf English Language stemmers are freely available) • Stop words to, the, it, be, or, … how about “to be or not to be” • Thesaurus fast -> rapid 15-853

  12. Other Methods • Document Ranking: • Returning an ordered ranking of the results • A priori ranking of documents (e.g. Google) • Ranking based on “closeness” to query • Ranking based on “relevance feedback” • Clustering and “Dimensionality Reduction” • Return results grouped into clusters • Return results even if query terms does not appear but are clustered with documents that do • Document Preprocessing • Removing near duplicates • Detecting spam 15-853

  13. Indexing and Searching Outline • Introduction: model, query types • Inverted File Indices: • Index compression • The lexicon • Merging terms (unions and intersections) • Vector Models: • Latent Semantic Indexing: • Link Analysis: PageRank (Google), HITS • Duplicate Removal: 15-853

  14. Documents as Bipartite Graph • Called an “Inverted File” index • Can be stored using adjacency lists, also called • posting lists (or files) • inverted file entry • Example size of TREC database(Text REtrieval Conference) • 538K terms • 742K documents • 333,856K edges • For the web, multiply by 10K Doc 1 Aardvark … … terms Documents 15-853

  15. Documents as Bipartite Graph • Implementation Issues: • 1. Space for posting lists these take almost all the space • 2. Access to lexicon • btrees, tries, hashing • prefix and wildcard queries • 3. Merging posting list • multiple term queries Lexicon Doc 1 Aardvark … … terms Documents 15-853

  16. 1. Space for Posting Lists • Posting lists can be as large as the document data • saving space and the time to access the space is critical for performance • We can compress the lists, • but, we need to uncompress on the fly. • Difference encoding: Lets say the term elephant appears in documents: • [3, 5, 20, 21, 23, 76, 77, 78] then the difference code is • [3, 2, 15, 1, 2, 53, 1, 1] 15-853

  17. Some Codes • Gamma code: if most significant bit of n is in location k, then gamma(n) = 0k n[k..0] 2 log(n) – 1 bits • Delta code: gamma(k)n[k..0] 2 log(log(n)) + log(n) - 1 bits • Frequency coded: base on actual probabilities of each distance 15-853

  18. Global vs. Local Probabilities • Global: • Count # of occurrence of each distance • Use Huffman or arithmetic code • Local: generate counts for each list elephant: [3, 2, 1, 2, 53, 1, 1] Problem: counts take too much space Solution: batching group into buckets by blog(length) c 15-853

  19. Performance • Bits per edge based on the TREC document collection • Total size = 333M * .66 bytes = 222Mbytes 15-853

  20. 2. Accessing the Lexicon • We all know how to store a dictionary, BUT… • it is best if lexicon fits in memory---can we avoid storing all characters of all words • what about prefix or wildcard queries? • Some possible data structures • Front Coding • Tries • Perfect Hashing • B-trees 15-853

  21. Front Coding • For large lexicons can save 75% of space • But what about random access? 15-853

  22. Prefix and Wildcard Queries • Prefix queries • Handled by all access methods except hashing • Wildcard queries • n-gram • rotated lexicon 15-853

  23. eb jezebel ec el … … n-grams terms n-gram • Consider every block of n characters in a term: e.g. 2-gram of jezebel -> $j,je,ez,ze,eb,el,l$ • Break wildcard query into an n-grams and search. • e.g. j*el would • search for $j,el,l$ as if searching for documents • find all potential terms • remove matches for which the order is incorrect 15-853

  24. Rotated Lexicon • Consider every rotation of a term: e.g. jezebel -> $jezebel, l$jezebe, el$jezeb, bel$jeze • Now store lexicon of all rotations • Given a query find longest contiguous block (with rotation) and search for it: e.g. j*el -> search for el$j in lexicon • Note that each lexicon entry corresponds to a single term e.g. ebel$jez can only mean jezebel 15-853

  25. 3. Merging Posting Lists • Lets say queries are expressions over: • and, or, andnot • View the list of documents for a term as a set: • Then e1 and e2 -> S1 intersect S2 e1 or e2 -> S1 union S2 e1 andnot e2 -> S1 diff S2 • Some notes: • the sets are ordered in the “posting lists” • S1 and S2 can differ in size substantially • might be good to keep intermediate results • persistence is important 15-853

  26. Union, Intersection, and Merging • Given two sets of length n and m how long does it take for intersection, union and set difference? • Assume elements are taken from a total order (<) • Very similar to merging two sets A and B, how long does this take? • What is a lower bound? 15-853

  27. Union, Intersection, and Merging • Lower Bound: • There are n elements of A and n + m positions in the output they could belong • Number of possible interleavings: • Assuming comparison based model, the decision tree has that many leaves • Max depth is at least log of number of leaves • Assuming m < n: 15-853

  28. Merging: Upper bounds • Brown and Tarjan show anO(m log((n + m)/m)) upper bound using 2-3 trees with cross links and parent pointers. Very messy. • We will take different approach, and base an implementation on two operations: split and join • Split and Join can then be implemented on many different kinds of trees. We will describe an implementation based on treaps. 15-853

  29. Split and Join • Split(S,v) : Split S into two sets S< = {s 2 S | s < v} and S> = {s 2 S | s > v}. Also return a flag which is true if v 2 S. • Split({7,9,15,18,22}, 18) ! {7,9,15},{22},True • Join(S<, S>) : Assuming 8 k<2 S<, k> in S> : k< < k>returns S<U S> • Join({7,9,11},{14,22}) ! {7,9,11,14,22} 15-853

  30. Time for Split and Join • Split(S,v)! (S<, S>),flag Join(S<, S>)! S • Naively: • T = O(|S|) • Less Naively: • T = O(log|S|) • What we want: • T = O(log(min(|S<|, |S>|))) -- can be shown • T = O(log |S<|) -- will actually suffice 15-853

  31. Will also use • isEmpty(S)! boolean • True if the set S is empty • first(S)! e • returns the least element of S • first({2,6,9,11,13}) ! 2 • {e}! S • creates a singleton set from an element • We assume they can both run in O(1) time. • An ADT with 5 operations! 15-853

  32. A a1 a2 a3 a4 a5 B b1 b2 b3 b4 b5 Out … b1 a1 b2 a2 b3 a3 b4 a4 Union with Split and Join • Union(S1, S2) = • if isEmpty(S1) thenreturn S2 • else • (S2<, S2>, fl) = Split(S2, first(S1)) • return Join(S2<, Union(S2>, S1)) 15-853

  33. Runtime of Union … • Tunion = O(åi log |oi| + åi log |oi|) • Splits Joins • Since the logarithm function is concave, this is maximized when blocks are as close as possible to equal size, therefore • Tunion = O(åi=1m log d n/m + 1 e) • = O(m log ((n+m)/m)) Out o1 o2 o3 o4 o5 o6 o7 o8 15-853

  34. Intersection with Split and Join • Intersect(S1, S2) = • if isempty(S1) thenreturnÆ • else • (S2<, S2>, flag) = Split(S2, first(S1)) • if flag then • return Join({first(S1)}, Intersect(S2>, S1)) • else • return Intersect(S2>, S1) 15-853

  35. Efficient Split and Join • Recall that we want: T = O(log |S<|) • How do we implement this efficiently? 15-853

  36. Treaps • Every key is given a “random” priority. • keys are stored in-order • priorities are stored in heap-order • e.g. (key,priority) : (1,23), (4,40), (5,11), (9,35), (12,30) (4,40) (1,23) (9,35) (5,11) (12,30) If the priorities are unique, the tree is unique. 15-853

  37. Left Spinal Treap • Time to split = length of path from Start to split location l • We will show that this is O(log L) in the expected case, where L is the number of keys between Start and l (inclusive). 10 in the example. • Time to Join is the same l Start 15-853

  38. Analysis Start l 15-853

  39. Analysis Continued • Proof: • i is an ancestor of j iff i has a greater priority than all elements between i and j, inclusive. • there are |i-j|+1 such elements each with equal probability of having the highest priority. Lemma: 15-853

  40. Analysis Continued • Can similarly show that: • Therefore the expected path length and runtime for split and join is O(log l). • Similar technique can be used for other properties of Treaps. 15-853

  41. And back to “Posting Lists” • We showed how to take Unions and Intersections, but Treaps are not very space efficient. • Idea: if priorities are in the range [0..1) then any node with priority < 1 - a is stored compressed. • a represents fraction of uncompressed nodes. 15-853

  42. Case Study: AltaVista • How AltaVista implements indexing and searching, or at least how they did in 1998. • Based on a talk by A. Broder and M. Henzinger from AltaVista. Henzinger is now at Google, Broder is at IBM. • The index (posting lists) • The lexicon • Query merging (or, and, andnot queries) • The size of their whole index is about 30% the size of the original documents it encodes. 15-853

  43. AltaVista: the index • All documents are concatenated together into one sequence of terms (stop words removed). • This allows proximity queries • Other companies do not do this, but do proximity tests in a postprocessing phase • Tokens separate documents • Posting lists contain pointers to individual terms in the single “concatenated” document. • Difference encoded • Use Front Coding for the Lexicon 15-853

  44. AltaVista: the lexicon • The Lexicon is front coded. • Allows prefix queries, but requires prefix to be at least 3 characters (otherwise too many hits) 15-853

  45. AltaVista: query merging • Support expressions on terms involving:AND, OR, ANDNOT and NEAR • Implement posting list with an abstract data type called an “Index Stream Reader” (ISR). • Supports the following operations: • loc() : current location in ISR • next() : advance to the next location • seek(k) : advance to first location past k 15-853

  46. AltaVista: query merging (cont.) • Queries are decomposed into the following operations: Create : term ! ISR ISR for the term Or: ISR * ISR ! ISR Union And: ISR * ISR ! ISR Intersection AndNot : ISR * ISR ! ISR Set difference Near: ISR * ISR ! ISR Intersection, almost • Note that all can be implemented with our Treap Data structure. • I believe (from private conversations) that they use a two level hierarchy that approximates the advantages of balanced trees (e.g. treaps). 15-853

More Related